---
title: "Demographic Structure and International Capital Flows: Evidence from 140 Countries"
author: "Working Paper"
date: "February 2026"
abstract: |
  We investigate the relationship between demographic structure and current account balances using the Fair-Dominguez polynomial age-distribution technique within the IMF's External Balance Assessment framework. Using data through 2024 from ten international sources across 140 countries---covering 97 percent of world population and 99 percent of world GDP---we estimate pooled GLS models with AR(1) correction. All three demographic polynomial variables are highly significant in the baseline model (all p < 0.001), with R-squared of 0.27 across 2,730 observations from 137 countries. However, a leave-one-region-out jackknife reveals moderate coefficient instability (CV approximately 30 percent), with the Central Asia and Caucasus group of 13 transition economies acting as the tipping point for statistical significance; these countries exhibit demographic slopes 3--8 times larger than the rest of the world, and direct testing rules out remittance dependence as the mechanism. Interactions between demographic polynomials and the Chinn-Ito financial openness index are jointly significant (p < 0.001), but a three-way interaction analysis reveals they operate exclusively through low- and middle-income countries (all p > 0.76 for high-income economies); a horse race against income-level interactions confirms this is genuinely about financial openness, not development level. In advanced economies, pension system generosity rather than financial openness mediates the demographic channel: demographic-pension interactions are jointly significant (p = 0.038 on EBA-49, all individual p < 0.008 on the full sample) and survive a horse race against KAOPEN interactions. The interest rate channel, insignificant through bank lending rates (p = 0.43), appears more clearly through government bond yields (p = 0.031) and term spreads (p < 0.001), though a two-stage test finds the demographics-to-yields first stage only suggestive (Z p-values 0.12--0.18). Projections through 2060 for 141 countries show that 92 economies currently face demographic headwinds, though the magnitude of projected effects scales linearly with the Z coefficients (CV approximately 30 percent across jackknife samples). A general equilibrium clearing rate overlay shows that the interest rate channel can absorb only 6--12 percent of peak GDP-weighted demographic surplus pressure, confirming that non-rate adjustment channels must bear the primary burden.
keywords: "demographics, current account, capital flows, age distribution, financial openness, pensions, lifecycle hypothesis"
jel: "F21, F32, J11, E21"
bibliography: references.bib
---

# Introduction

The world is aging unevenly. Japan's median age has surpassed 48 years while Nigeria's remains below 18. China's working-age population is contracting by millions annually following decades of the one-child policy, even as India adds young workers at record pace. These divergent demographic trajectories have profound implications for savings, investment, and the international allocation of capital.

The lifecycle hypothesis [@modigliani1954] provides the foundational mechanism: young dependents consume without producing, working-age adults accumulate savings, and retirees draw down their assets. At the national level, these individual decisions aggregate into macroeconomic patterns---countries with large working-age cohorts tend to save more than they invest domestically, generating current account surpluses and capital outflows, while countries with dependent-heavy populations tend to run deficits.

This insight has been formalized through polynomial representations of age distributions. @fair1991 introduced the technique of constraining age-group coefficients via a low-order polynomial, reducing the dimensionality problem inherent in estimating separate effects for each of 17 age groups. @higgins1998 applied this approach to an international panel, demonstrating that demographics significantly predict savings, investment, and current account balances. @arnott2012 extended the technique to financial asset returns. Most recently, @koomen2020 embedded the Fair-Dominguez polynomial in the IMF's External Balance Assessment (EBA) model, which represents the state of the art in multilateral current account modeling [@imfeba2013; @imfeba2019].

This paper makes seven contributions. First, we embed the Koomen-Wicht framework in a substantially broader sample with data through 2024, incorporating the UN World Population Prospects 2024 revision and the latest IMF World Economic Outlook projections. Second, we assemble the broadest country sample in the demographic-current account literature: 140 countries covering 97 percent of world population and 99 percent of GDP, far exceeding the standard EBA 49. Third, we introduce interactions between demographic polynomial variables and the Chinn-Ito capital openness index (KAOPEN), testing whether financial integration gates the demographic channel---a hypothesis rooted in @higgins1998 and developed theoretically by @carvalho2016. A three-way interaction analysis (demographics $\times$ KAOPEN $\times$ income group) reveals that financial openness matters exclusively for developing countries, not advanced economies; a horse race against income-level interactions confirms this is genuinely about openness, not development level. Fourth, we identify pension system generosity as the robust institutional mediator in advanced economies: demographic-pension interactions are significant (p = 0.038), survive a horse race against KAOPEN interactions, and strengthen on the full sample (all p < 0.008). Fifth, we resolve the interest rate puzzle by showing that the rate channel operates through government bond yields and term spreads rather than bank lending rates, with the demographic signal concentrated in the yield curve slope. Sixth, we provide projections through 2060 for 141 countries with a general equilibrium clearing rate overlay. Seventh, we conduct a thorough coefficient stability analysis---leave-one-region-out jackknife, cumulative sample build-up, and extended model sensitivity---that documents the dependence of baseline results on specific country groups and provides a transparent assessment of the fragility of demographic-current account estimates.

Our results establish individual significance of all three demographic polynomial variables (all p < 0.001) in the baseline model---a result enabled by the breadth of our country sample. The demographic polynomials and EBA controls achieve an R-squared of 0.27 across 137 countries. The extended model with KAOPEN interaction terms reaches 0.29 across 90 countries, with all interactions significant (all p < 0.05) and jointly highly significant (p < 0.001). However, a leave-one-region-out jackknife reveals that these coefficients are moderately unstable (CV approximately 30 percent), with the Central Asia and Caucasus group of 13 countries acting as the critical region for significance. This sensitivity---documented transparently and in detail---suggests that the broader literature's difficulty in establishing robust demographic-current account estimates reflects genuine identification challenges inherent in pooled cross-country analysis.

A central finding is that the institutional channel through which demographics affect current accounts differs by income level. Among developing countries, financial openness gates the demographic channel: closed economies with young populations see no demographic-CA effect, while similar economies that have opened show the lifecycle pattern. Among advanced economies---already at the KAOPEN ceiling---it is pension system generosity that determines the strength of the demographic-CA relationship. This dual-channel result reconciles conflicting findings in the prior literature, which has variously emphasized openness [@higgins1998] and pension systems [@boersch2010] without recognizing that the relevant mechanism depends on the country's stage of development.

A systematic heterogeneity analysis reveals that the shape of the implied age-coefficient profile varies substantially across country groups. Formal interaction models identify savings behavior (all p < 0.002) and income level (all p < 0.008) as the dominant mediating channels, while trade openness and gross financial integration are not significant. This is consistent with the demographic effect operating through lifecycle savings mechanisms rather than trade or financial volume channels.

Projections through 2060 for 141 countries reveal that 92 economies currently face demographic headwinds on their current accounts, with only 30 receiving tailwinds. The largest future swings are in East Asia (Hong Kong +20.9 pp, Korea +16.3 pp, Taiwan +15.2 pp) and Sub-Saharan Africa and Central America (Lesotho -7.3 pp, Guatemala -6.7 pp). A general equilibrium clearing rate overlay shows that with near-universal coverage, the interest rate channel can absorb only 6--12 percent of peak GDP-weighted demographic surplus pressure---far less than suggested by studies limited to advanced economies---confirming that exchange rate adjustment, fiscal policy, and structural reform must bear the primary equilibrating burden.

The remainder of this paper is organized as follows. Section 2 reviews the related literature. Section 3 presents the methodology. Section 4 describes the data. Section 5 presents the main results. Section 6 examines structural stability and coefficient sensitivity. Section 7 provides demographic projections. Section 8 discusses implications and limitations. Section 9 concludes.


# Literature Review

## The Fair-Dominguez Polynomial Technique

@fair1991 introduced a method for incorporating age-distribution effects into macroeconomic models without estimating separate coefficients for each age group. With $G$ age groups, direct estimation would require $G$ parameters, leading to severe multicollinearity given that population shares sum to one. Their solution constrains the age-group coefficients $\alpha_g$ to lie on a polynomial of degree $P$:

$$\alpha_g = \sum_{p=0}^{P} \gamma_p g^p$$

where $g = 1, \ldots, G$ indexes the age groups. This reduces the parameter space from $G$ to $P+1$, and the individual $\alpha_g$ can be recovered from the estimated $\gamma_p$ for interpretation.

## Demographics and Current Accounts

@higgins1998 was the first to apply this technique to an international panel of savings, investment, and current account balances. Using a sample of over 100 countries from 1950-1992, Higgins demonstrated that demographic variables explain a substantial share of cross-country variation in external balances. A key finding was that demographic effects on current accounts are significant primarily in financially open economies---a result we revisit with updated data and the KAOPEN index, finding that it holds for developing but not advanced economies.

@koomen2020 advanced this literature by embedding the demographic polynomial in the IMF's External Balance Assessment framework. The EBA model [@imfeba2013; @imfeba2019] is the standard tool for multilateral current account assessment, incorporating institutional, policy, and cyclical determinants alongside demographics. Koomen and Wicht showed that the demographic polynomial adds significant explanatory power beyond standard EBA controls, and that the implied age-group coefficients display the theoretically expected lifecycle pattern. Their sample was limited to 49 countries; we extend this to 140.

## Interest Rates and Demographics

A complementary literature examines how demographics affect the natural rate of interest and thereby influence capital flows indirectly. @carvalho2016 develop a theoretical model in which aging depresses the equilibrium real interest rate by shifting savings supply. @rachel2017 attribute a significant portion of the secular decline in global real rates to demographic factors. @summers2014 frames this as "secular stagnation"---the possibility that aging advanced economies face chronically insufficient demand. The link to capital flows is direct: if demographics compress interest rates in aging countries, capital should flow outward toward younger, higher-return economies. @bernanke2005 identified the "global saving glut"---excess savings from aging East Asian and oil-exporting economies depressing US interest rates and financing the US current account deficit---as a fundamentally demographic phenomenon.

## The Lucas Paradox

@lucas1990 famously observed that capital does not flow from rich to poor countries as standard neoclassical theory predicts. Demographics offer a partial resolution: if rich countries are also aging countries, their excess savings may flow to other aging but financially open economies (like the United States) rather than to young but financially closed developing economies. Our analysis of KAOPEN interactions speaks directly to this puzzle---and our finding that the interaction operates through developing countries suggests that the Lucas paradox is partly a demographic gating problem: capital does not flow to young countries because their capital accounts are closed.


# Methodology

## Age-Share Construction

We define $G = 17$ five-year age groups (0-4, 5-9, $\ldots$, 75-79, 80+), indexed $g = 1, \ldots, 17$. For each country $i$ and year $t$, we compute the population share in group $g$:

$$n_{g,it} = \frac{N_{g,it}}{N_{it}}$$

where $N_{g,it}$ is the population in age group $g$ and $N_{it}$ is total population. Following the EBA methodology, we express shares as deviations from the GDP-weighted world average:

$$\tilde{n}_{g,it} = n_{g,it} - \sum_j w_{jt} n_{g,jt}$$

where $w_{jt}$ is country $j$'s share of world GDP in year $t$. This demeaning ensures that the demographic variables capture country-specific departures from the global age structure, consistent with the EBA's focus on explaining cross-country differences.

## Polynomial Constraint

We impose a cubic polynomial constraint ($P = 3$) on the age-group coefficients:

$$\alpha_g = \gamma_0 + \gamma_1 g + \gamma_2 g^2 + \gamma_3 g^3$$

with the zero-sum restriction $\sum_{g=1}^{G} \alpha_g = 0$, which ensures that a uniform age distribution has no net effect on the current account. This restriction pins down $\gamma_0$ as a function of $\gamma_1, \gamma_2, \gamma_3$:

$$\gamma_0 = -\frac{1}{G} \sum_{g=1}^{G} \left( \gamma_1 g + \gamma_2 g^2 + \gamma_3 g^3 \right)$$

The three estimable parameters $\gamma_1, \gamma_2, \gamma_3$ enter the regression through transformed demographic variables $Z_1, Z_2, Z_3$, constructed from the demeaned age shares:

$$Z_{p,it} = \sum_{g=1}^{G} g^p \cdot \tilde{n}_{g,it}, \quad p = 1, 2, 3$$

## Model Specification

The baseline model takes the form:

$$\frac{CA_{it}}{GDP_{it}} = \gamma_1 Z_{1,it} + \gamma_2 Z_{2,it} + \gamma_3 Z_{3,it} + \boldsymbol{\beta}' \mathbf{X}_{it} + u_{it}$$

where $\mathbf{X}_{it}$ is a vector of EBA control variables including fiscal balance/GDP, the Chinn-Ito financial openness index (KAOPEN), expected GDP growth, lagged net foreign assets/GDP, log relative output per worker, and life expectancy.

The extended model adds an interest rate variable (log lending rate) and interactions between the demographic polynomials and financial openness:

$$\frac{CA_{it}}{GDP_{it}} = \gamma_1 Z_{1,it} + \gamma_2 Z_{2,it} + \gamma_3 Z_{3,it} + \boldsymbol{\beta}' \mathbf{X}_{it} + \sum_{p=1}^{3} \delta_p (Z_{p,it} \times KAOPEN_{it}) + u_{it}$$

These interaction terms test whether demographic effects on current accounts are amplified or attenuated by the degree of capital account openness.

## Estimation

Following EBA practice, we estimate pooled GLS without country fixed effects. The error term follows a panel-wide AR(1) process:

$$u_{it} = \rho \cdot u_{i,t-1} + \varepsilon_{it}$$

We estimate $\rho$ and the model parameters jointly using iterative Cochrane-Orcutt with a Prais-Winsten transformation for the first observation of each country panel. This procedure accounts for the strong serial correlation typical of annual current account data while preserving the cross-sectional variation that identifies demographic effects.

## Recovery of Age-Group Coefficients

After estimation, we recover the implied age-group coefficients $\hat{\alpha}_g$ from the estimated polynomial parameters $\hat{\gamma}_0, \hat{\gamma}_1, \hat{\gamma}_2, \hat{\gamma}_3$. These recovered coefficients can be plotted against age groups to visualize the lifecycle pattern of demographic effects on the current account.


# Data

## Sources

We assemble a comprehensive panel dataset from ten international sources, summarized in Table 1.

: Data Sources and Coverage {#tbl:sources}

![](tables/data_sources.md)

The primary demographic data come from the United Nations World Population Prospects 2024 revision [@unwpp2024], which provides population estimates by five-year age group for 237 countries and territories from 1950 to 2100. We use estimates through 2024 and medium-variant projections from 2025 onward.

Macroeconomic fundamentals are drawn from the IMF World Economic Outlook (April 2025 vintage) for current account balances, fiscal balances, GDP growth, and output gaps. The Penn World Tables version 10.0 [@pwt100] provide output per worker and PPP-adjusted GDP. The World Bank World Development Indicators supply health expenditure and life expectancy data. Financial openness is measured by the Chinn-Ito KAOPEN index [@chinn2006], and net foreign asset positions come from the Lane and Milesi-Ferretti External Wealth of Nations dataset [@lane2007].

Interest rate data combine two sources: FRED provides OECD government bond yields and short-term rates for 23 advanced economies, while the IMF Monetary and Financial Statistics database supplies lending and money market rates for a broader sample of 132 countries.

## Country Sample

Our estimation sample encompasses 140 countries organized in four groups. The core consists of the 49 countries in the IMF's External Balance Assessment. We extend this with 20 Sub-Saharan African economies to capture the young-population tail of the demographic distribution. We then add 10 EU member states not in the EBA-49 (Romania, Slovakia, Bulgaria, Croatia, Lithuania, Slovenia, Estonia, Latvia, Cyprus, Malta) to achieve near-complete EU coverage. Finally, we include 61 additional economies across Asia (Bangladesh, Vietnam, Cambodia, Iran, Kazakhstan, and others), the Middle East and Central Asia (Iraq, Qatar, Kuwait, Oman, Uzbekistan), Latin America and the Caribbean (Dominican Republic, Ecuador, Guatemala, Costa Rica, Jamaica), Eastern Europe and the Caucasus (Ukraine, Belarus, Georgia, Armenia, Albania), and additional Sub-Saharan African economies. The complete country list is provided in Appendix B.

The effective sample depends on data availability across all required variables. The demographics-only model uses 5,323 observations from 141 countries. The baseline model, which requires EBA control variables, uses 2,730 observations from 137 countries covering 1986--2024. The extended model, which additionally requires lending rates, uses 1,626 observations from 90 countries. The sample covers 97 percent of world population and 99 percent of world GDP.

## Summary Statistics

: Summary Statistics {#tbl:sumstats}

![](tables/summary_statistics.md)

Table 2 reports summary statistics for the key variables in the estimation sample. The mean current account balance is -1.9% of GDP, reflecting the preponderance of deficit countries in the sample. The demographic polynomial variables ($Z_1$, $Z_2$, $Z_3$) display substantial cross-country variation, driven by the wide range of age structures from young Sub-Saharan African populations to aging East Asian and European societies.


# Results

## Model Specifications

We estimate four specifications of increasing complexity. Model 1 includes only the three demographic polynomial variables. Model 2 adds EBA control variables, selected through a stepwise procedure that begins with core controls (fiscal balance, KAOPEN, expected growth) and adds secondary controls (lagged NFA/GDP, relative output per worker, health expenditure, life expectancy) only if they do not reduce the estimation sample by more than 30%. Fiscal balance/GDP is winsorized at the 1st and 99th percentiles to limit the influence of extreme outliers (the raw variable ranges from -557% to +125% of GDP due to hyperinflation and conflict episodes). Model 3 further adds the log lending rate and demographic-KAOPEN interaction terms. The lending rate enters as log(1 + rate/100) rather than in levels, because raw lending rates range from 0% to nearly 100,000% due to hyperinflation episodes; the log transformation maps rates to a continuously compounded scale. Model 4 restricts the sample to countries with pension data and adds old-age public spending and demographic-pension interactions.

: Regression Results {#tbl:regression}

| Variable | Model 1 | Model 2 | Model 3 | Model 4 |
|:---------|---:|---:|---:|---:|
| Z_1 | 38.4*** | 48.3*** | 25.0 | -407.7*** |
| | (10.2) | (14.5) | (20.6) | (128.4) |
| Z_2 | -5.0*** | -7.9*** | -4.3 | 55.4*** |
| | (1.5) | (2.2) | (3.1) | (17.7) |
| Z_3 | 0.19*** | 0.34*** | 0.19 | -2.11*** |
| | (0.06) | (0.09) | (0.13) | (0.68) |
| fiscal_bal_gdp | | 0.31*** | 0.38*** | -0.06 |
| | | (0.03) | (0.04) | (0.06) |
| kaopen | | -0.82*** | -0.67 | -0.07 |
| | | (0.22) | (0.42) | (0.35) |
| expected_growth | | -0.23*** | -0.44*** | -0.10 |
| | | (0.06) | (0.10) | (0.10) |
| nfa_gdp_lag | | 0.62*** | 0.51** | -0.22 |
| | | (0.21) | (0.25) | (0.34) |
| log_rel_opw | | 3.99*** | 5.23*** | 2.87** |
| | | (0.54) | (0.77) | (1.42) |
| health_exp_gdp | | -0.70*** | -0.85*** | -0.74** |
| | | (0.23) | (0.29) | (0.35) |
| log_lending_rate | | | 3.68 | |
| | | | (4.71) | |
| pension_spending_gdp | | | | 0.25** |
| | | | | (0.12) |
| Z_1 x KAOPEN | | | 24.7** | |
| | | | (11.9) | |
| Z_2 x KAOPEN | | | -4.2** | |
| | | | (1.8) | |
| Z_3 x KAOPEN | | | 0.18** | |
| | | | (0.07) | |
| Z_1 x pension | | | | 16.6*** |
| | | | | (6.2) |
| Z_2 x pension | | | | -2.3*** |
| | | | | (0.8) |
| Z_3 x pension | | | | 0.09*** |
| | | | | (0.03) |
| R-squared | 0.056 | 0.273 | 0.290 | 0.137 |
| N (countries) | 5,323 (141) | 2,730 (137) | 1,626 (90) | 750 (40) |
| AR(1) rho | 0.652 | 0.812 | 0.808 | 0.825 |

*Notes: Pooled GLS with iterative Cochrane-Orcutt AR(1) correction. Standard errors in parentheses. Significance: * p<0.05, ** p<0.01, *** p<0.001.*

In Model 1 (demographics only), all three polynomial variables are individually highly significant (all p < 0.001) across 5,323 observations from 141 countries, explaining 5.6% of current account variation. The relatively low R-squared reflects the broad heterogeneity of the 141-country sample; the AR(1) coefficient of 0.65 indicates moderate persistence.

Model 2 (baseline with EBA controls) achieves R-squared = 0.27 across 2,730 observations from 137 countries. The demographic polynomials remain highly significant: Z_1 = 48.3 (p = 0.001), Z_2 = -7.9 (p < 0.001), Z_3 = 0.34 (p < 0.001). Fiscal balance is the dominant control, with a coefficient of 0.31 (p < 0.001), implying that a one percentage point improvement in the fiscal balance is associated with a 0.31 percentage point increase in the current account. KAOPEN enters negatively and highly significantly (-0.82, p < 0.001), indicating that more financially open countries tend to run lower current account surpluses on average---consistent with openness facilitating capital inflows in the many developing economies in our sample. Relative output per worker is large and highly significant (3.99, p < 0.001), reflecting the Balassa-Samuelson effect visible across the wide productivity range of 140 countries. NFA persistence is significant (0.62, p = 0.003), capturing the feedback from accumulated external positions to current flows.

Model 3 (extended with interactions) confirms the significance of demographic-KAOPEN interactions. Z_1 $\times$ KAOPEN enters at 24.7 (p = 0.039), Z_2 $\times$ KAOPEN at -4.2 (p = 0.021), and Z_3 $\times$ KAOPEN at 0.18 (p = 0.013). The joint F-test is highly significant (p < 0.001). In the extended model, the direct demographic coefficients represent the effect at zero financial openness, while the total effect for a financially open economy is the sum of direct and interaction terms. The log lending rate is not significant (p = 0.43), indicating that the interest rate channel does not operate through bank lending rates---a finding we investigate further using alternative rate measures below.

Model 4 (pension interactions) is estimated on the 40-country subsample with pension spending data and reveals that pension generosity is a powerful mediator of the demographic channel. All three Z $\times$ pension interactions are individually highly significant (all p < 0.008), and the pension level itself is significant (0.25, p = 0.03). The direct demographic coefficients reverse sign and become large when pension interactions are included, indicating that the pension system absorbs much of the lifecycle mechanism in the countries where it is measured. This model is discussed further in Section 5.3.

## Implied Age-Group Coefficients

Figure 1 displays the implied age-group coefficients recovered from the baseline model's polynomial estimates. The pattern conforms to lifecycle theory: young age groups (0-14) exert negative pressure on the current account (consumption without production), working-age adults in their 30s and 40s exert the strongest positive pressure (peak saving years), and elderly groups return toward negative territory (retirement dissaving).

![Implied age-group coefficients from the cubic polynomial. Red bars indicate negative effects (deficit pressure), green bars indicate positive effects (surplus pressure). Error bars show 95% confidence intervals.](figures/fig1_age_coefficients.png){#fig:agecoefs}

The peak positive coefficient occurs at ages 35-39, with the characteristic lifecycle hump shape. This is younger than the 55-59 peak reported by @koomen2020, a difference we attribute to our broader country sample and investigate through heterogeneity analysis below.

## Institutional Mediators: Financial Openness and Pensions

### Financial Openness Operates Through Developing Countries

The significance of the KAOPEN interactions in Model 3 raises the question: does financial openness amplify the demographic channel universally, or only in certain country groups? We test this by estimating a three-way interaction model, classifying countries into three income groups by GDP per capita PPP terciles and allowing the Z $\times$ KAOPEN interactions to differ across groups.

: Three-Way Interaction: Z x KAOPEN x Income Group {#tbl:threeway}

| Variable | High Income | Middle Income | Low Income |
|:---------|---:|---:|---:|
| Z_1 x K x income | 0.3 (p=0.99) | 24.6 (p=0.22) | 47.7 (p=0.07) |
| Z_2 x K x income | 0.5 (p=0.86) | -5.5 (p=0.06) | -7.5 (p=0.07) |
| Z_3 x K x income | -0.04 (p=0.76) | 0.28 (p=0.02) | 0.31 (p=0.07) |
| N countries | 21 | 35 | 34 |

*Notes: Three-way model R-squared = 0.355 vs 0.290 for two-way. Joint F-test for three-way vs two-way: F(6,1606) = 27.07, p < 0.001.*

The result is unambiguous: the Z $\times$ KAOPEN interaction is zero among high-income countries. All coefficients are near zero with p-values exceeding 0.76. The interaction operates through low- and middle-income countries, where the contrast between closed economies (low KAOPEN, young demographics) and opening economies identifies the effect. In a full specification with Z $\times$ income main effects, the Z $\times$ low-income interactions are all highly significant (all p < 0.007) while Z $\times$ high-income interactions are insignificant (all p > 0.79), confirming that demographics affect current accounts differently for low-income countries even before considering openness.

A natural concern is whether the KAOPEN interaction proxies for income level rather than financial openness per se, since the two are correlated (r = 0.33 in the developing subsample). We test this directly with a horse race: in the developing-country subsample (90 countries, 1,824 observations), we estimate a model with both Z $\times$ KAOPEN and Z $\times$ log(GDP per capita) interactions. The KAOPEN interactions survive (Z_2 $\times$ KAOPEN p = 0.013, Z_3 $\times$ KAOPEN p = 0.005) while the GDP interactions collapse (all p > 0.49). Joint F-tests confirm both sets are jointly significant (both p < 0.001), but individually KAOPEN dominates. The interaction is genuinely about financial openness, not a proxy for development level.

This reframes the @higgins1998 financial openness result. The mechanism is not that financial openness "gates" the demographic channel in advanced economies---they are already at the gate. Rather, the interaction captures the transition from financial autarky to integration in developing countries: as economies open their capital accounts, the lifecycle pattern begins to manifest in their current accounts.

### Pension Generosity Is the Advanced-Economy Channel

If financial openness is not the relevant institutional mediator for advanced economies, what is? @koomen2020 included old-age dependency and social spending controls in their EBA-based specification; we extend this by directly interacting pension generosity with the demographic polynomials to test whether pension systems amplify the lifecycle mechanism. We test pension system generosity using OECD Social Expenditure data on old-age public spending as a share of GDP.

: Pension Model Horse Race on EBA-49 Subsample {#tbl:pension}

| Variable | + Pension | + Z x Pension | Horse Race (Both) |
|:---------|---:|---:|---:|
| pension_spending_gdp | 0.26* | 0.21 | 0.22 |
| Z_1 x pension | --- | 10.4 (p=0.09) | 11.6 (p=0.06) |
| Z_2 x pension | --- | -1.5 (p=0.07) | -1.6 (p=0.05) |
| Z_3 x pension | --- | 0.06 (p=0.06) | 0.06 (p=0.04)* |
| Z_1 x KAOPEN | --- | --- | -32.2 (p=0.47) |
| Z_2 x KAOPEN | --- | --- | 4.0 (p=0.51) |
| Z_3 x KAOPEN | --- | --- | -0.13 (p=0.54) |
| Joint F (Z x pension) | --- | p=0.038 | Pension survives |
| N (countries) | 597 (31) | 597 (31) | 597 (31) |

The Z $\times$ pension interactions are jointly significant (F(3,583) = 2.83, p = 0.038) in the EBA-49 subsample. In the horse race model that includes both pension and KAOPEN interactions simultaneously, the pension interactions survive (Z_3 $\times$ pension individually significant at p = 0.043) while all KAOPEN interactions are completely insignificant (all p > 0.47). On the full 140-country sample (40 countries with pension data, Model 4 in Table 3), all three Z $\times$ pension interactions become individually highly significant (all p < 0.008), with R-squared of 0.14.

The interpretation is that in advanced economies, pension system generosity---not financial openness---determines how strongly demographics translate to current account positions. More generous pension systems amplify the lifecycle mechanism by strengthening the link between age structure and aggregate savings behavior. Financial openness, at the ceiling for virtually all advanced economies, provides no additional variation.

## Heterogeneity of the Age-Coefficient Profile

### Demographic Contributions Across Country Groups

Rather than re-estimating the polynomial on subsamples (where the three-parameter Fair-Dominguez specification can become unstable with limited demographic variation---see Appendix D), we apply the full-sample baseline coefficients to each country group's demographic data. This reveals how the *same* lifecycle mechanism produces different current account effects across countries with different age structures.

: Projected Demographic Contributions by Country Group (Full-Sample Coefficients) {#tbl:subsamples}

| Country Group | N countries | N obs | Mean Demo Contribution (pp) | Mean Actual CA/GDP (pp) |
|:--|--:|--:|--:|--:|
| Full Sample | 137 | 2,730 | -1.8 | -1.2 |
| EBA-49 only | 49 | 1,378 | -0.1 | +0.8 |
| OECD | 21 | 669 | +0.2 | +0.5 |
| Emerging (non-OECD/SSA) | 27 | 709 | -0.4 | +1.2 |
| SSA | 37 | 815 | -5.3 | -4.2 |
| High income | 46 | 932 | +0.4 | +2.7 |
| Middle income | 49 | 980 | -0.5 | -1.2 |
| Low income | 46 | 818 | -4.7 | -3.6 |

The lifecycle mechanism generates starkly different demographic pressures across country groups. OECD and high-income countries, with large working-age cohorts relative to dependents, receive a modest positive demographic contribution (+0.2 to +0.4 pp). Sub-Saharan African and low-income countries, burdened by youth dependency, face a large negative demographic drag (-4.7 to -5.3 pp). Emerging markets sit between these extremes (-0.4 pp), with many countries transitioning toward the demographic window where the contribution turns positive.

### Demographic Interaction Models

To test formally whether country characteristics mediate the demographic channel, we augment the baseline model with interaction terms between the demographic polynomials and six country characteristics:

$$\frac{CA_{it}}{GDP_{it}} = \gamma_1 Z_{1,it} + \gamma_2 Z_{2,it} + \gamma_3 Z_{3,it} + \boldsymbol{\beta}' \mathbf{X}_{it} + \sum_{p=1}^{3} \delta_p (Z_{p,it} \times W_{it}) + u_{it}$$

where $W_{it}$ is the mediating variable. Each interaction model is estimated separately to avoid multicollinearity.

: Demographic Interaction Model Results (Heterogeneity Sweep) {#tbl:interactions}

| Channel ($W$) | Z_1 x W (p) | Z_2 x W (p) | Z_3 x W (p) | R-squared | N |
|:--|:--|:--|:--|--:|--:|
| Savings/GDP | 338.6 (<0.001)*** | -43.4 (0.002)*** | 1.54 (0.006)*** | 0.446 | 1,848 |
| log(GDP/cap) | 36.1 (0.008)*** | -5.8 (0.007)*** | 0.25 (0.006)*** | 0.319 | 1,946 |
| KAOPEN | 16.0 (0.055)* | -2.7 (0.034)** | 0.12 (0.023)** | 0.314 | 1,946 |
| Gross IFI | -1.6 (0.542) | 0.3 (0.376) | -0.02 (0.284) | 0.385 | 1,946 |
| NFA/GDP | 17.7 (0.131) | -2.7 (0.136) | 0.11 (0.146) | 0.279 | 1,946 |
| Trade openness | 16.2 (0.546) | -1.5 (0.708) | 0.04 (0.824) | 0.338 | 1,858 |

Two channels dominate. **Savings/GDP** is the strongest mediator: all three interaction terms are significant at the 1% level, and the model achieves the highest R-squared (0.446) of any specification. **Income level** is the second-strongest channel, with all three interactions significant at the 1% level. Notably, **gross financial integration**, **net foreign asset position**, and **trade openness** do not significantly mediate the demographic channel. The demographic effect is statistically associated with savings behavior and de jure financial openness, not with the volume of cross-border financial positions or trade flows.

## Savings and Investment Channels

The lifecycle mechanism can operate through savings (demographics $\rightarrow$ national savings $\rightarrow$ CA) or investment (demographics $\rightarrow$ domestic investment $\rightarrow$ CA), or both. We test this by replacing the dependent variable with gross savings/GDP and gross investment/GDP respectively, holding the same controls.

Demographics add 7.7 percentage points of incremental R-squared to a savings equation and 10.6 percentage points to an investment equation. Both channels are active in the 140-country sample, with the investment channel slightly dominant. All three Z terms are individually significant in both savings and investment equations, confirming that age structure is associated with both sides of the current account identity.

## Nonlinearity Tests

Two control variables have strong theoretical priors for nonlinear effects: net foreign assets (creditor-debtor asymmetry) and life expectancy (convexity of the retirement dissaving effect).

### Net Foreign Asset Position

Adding NFA/GDP squared to the baseline specification yields a significant positive coefficient (0.041, p = 0.033), confirming convexity---the effect of NFA on the current account strengthens at extreme positions. A piecewise linear split reveals asymmetry: the coefficient for creditor nations (NFA > 0) is 0.40 (p = 0.08), while the coefficient for debtor nations (NFA < 0) is -0.54 and not significant. The creditor effect is marginally significant, indicating that countries with large positive NFA positions experience self-reinforcing surplus dynamics through the investment income channel, though the effect is less sharply identified than in narrower samples.

### Life Expectancy

Adding life expectancy squared produces a marginally significant quadratic term (0.008, p = 0.059), suggesting convexity in the life expectancy-current account relationship. Combined with the linear term, this implies a turning point at approximately 65 years.

### Joint Specification

Both nonlinear terms retain their significance when included simultaneously (NFA-squared p = 0.031, life expectancy-squared p = 0.056), with a joint R-squared of 0.39.

## Multiple Testing Correction

This paper conducts 26 distinct hypothesis tests across confirmatory and exploratory specifications. To guard against false discovery, we apply three standard corrections: Bonferroni (family-wise error rate control, most conservative), Holm step-down (uniformly more powerful than Bonferroni), and Benjamini-Hochberg (false discovery rate control at q = 0.05).

: Multiple Testing Results {#tbl:multtest}

| Category | Raw p < 0.05 | Bonferroni | Holm | BH (FDR) |
|:---------|---:|---:|---:|---:|
| All 26 tests | 22 | 11 | 12 | 21 |
| 7 confirmatory tests | 6 | 4 | 4 | 6 |

Among the 7 confirmatory tests (pre-specified replications of the demographic-current account model and KAOPEN interactions), the three baseline Z coefficients and the joint Z $\times$ KAOPEN F-test survive Bonferroni correction at the family level (threshold p < 0.007). The three individual Z $\times$ KAOPEN terms (p = 0.013--0.039) do not survive family-wise correction but do survive FDR correction.

Among exploratory tests, the following survive Bonferroni: the income-group joint F-tests (both middle and low income), the term spread, the savings and investment channels, and both sides of the KAOPEN vs. GDP horse race. The pension interactions on the full 140-country sample survive Holm correction (Z_3 $\times$ pension p = 0.003). The pension horse race on the EBA-49 subsample (Z_3 $\times$ pension p = 0.043) does not survive any correction and should be considered suggestive.

One test that was significant at raw p < 0.05 does not survive any correction: the Z_3 $\times$ pension horse race result (p = 0.043). We flag this explicitly as an exploratory finding requiring confirmation.

Throughout the remainder of this paper, we distinguish between findings that survive multiple testing correction (marked with "confirmed") and those that are exploratory and require replication (marked with "exploratory"). Readers should weight conclusions accordingly.

## Alternative Interest Rate Measures

The insignificance of the log lending rate in Model 3 (p = 0.43) may reflect the inadequacy of bank lending rates as a proxy for the capital-flow-relevant cost of capital. Lending rates embed bank margins, credit risk premia, and regulatory markups that obscure the underlying equilibrium rate. To test this, we re-estimate the extended model with two alternative measures: real 10-year government bond yield differentials and the term spread. Both are available for 23 OECD countries from FRED.

### Bond Yield Differentials (Model 3b)

Replacing log lending rate with the real bond yield differential produces a significant coefficient (0.13, p = 0.031) across 679 observations from 23 countries. The interest rate channel that was invisible through lending rates appears more clearly through bond-market measures.

### Term Spread (Model 3c)

The term spread is even more significant (0.23, p < 0.001). Countries with steeper yield curves run larger current account surpluses, consistent with the interpretation that a compressed term spread reflects expectations of persistently low future rates that push savings abroad.

### Decomposing the Yield Curve

The real short-rate differential is not significant (0.03, p = 0.59), meaning the demographic signal resides entirely in the *slope* of the yield curve, not the level of short rates. This is consistent with the lifecycle mechanism operating through long-horizon savings and investment decisions rather than through monetary policy transmission.

### Carry Trade Variables

The FX-hedged yield versus Japan is highly significant (0.31, p < 0.001), as is the unhedged carry versus Japan (-0.44, p = 0.001). Neither measure relative to the US is significant. Japan-specific carry trade variables outperform all other rate measures except the term spread, consistent with Japan's role as the world's primary carry trade originator.

### Two-Stage Carvalho Channel Test

In Stage 1, demographics do not individually predict real bond yield differentials (Z p-values: 0.12, 0.14, 0.18), though the signs are correct. In Stage 2, fitted bond yield differentials significantly predict current account balances (coefficient -3.4, p = 0.019). The negative coefficient implies that countries whose bond yields are low *for demographic reasons* tend to run surpluses---consistent with the @carvalho2016 mechanism where demographic savings pressure compresses rates and pushes capital abroad.


# Structural Stability

## Rolling-Window Estimation

We re-estimate the baseline model over 15-year rolling windows, advancing one year at a time from 1986-2000 through 2010-2024. We do not find statistically significant parameter breaks at major global dates, but the incremental explanatory power of demographics varies over time. The incremental R-squared from demographics peaked at 12.6% around 1999-2008, declined post-GFC, and settled at 2-4%. Critically, the Z_1 coefficient becomes individually significant from 2006 onward and remains so through the most recent windows, in contrast to narrower samples where significance fades post-GFC.

![Rolling-window estimates of Z_1, Z_2, Z_3 coefficients and R-squared over 15-year windows. Shaded bands show 95% confidence intervals. Vertical dashed lines mark WTO accession (2001), GFC (2008), and tariff onset (2018).](figures/fig3_rolling_coefficients.png){#fig:rolling}

The stability of the demographic signal in the broad sample, compared to its apparent weakening in narrower samples, is important. Prior studies documenting a post-GFC decline in the demographic-current account relationship were likely observing a sample composition effect: the demographic variation that drives identification was concentrated in the EBA-49 countries whose age structures converged during the 2000s. Adding the broader set of young-population and rapidly-aging economies provides sufficient variation for demographic effects to remain identifiable throughout the sample period.

## Split-Sample Structural Break Tests

: Split-Sample Structural Break Comparison {#tbl:breaks}

![](tables/structural_breaks.md)

Formal interaction tests (adding $Z \times post\_break$ terms to the full sample) yield no significant coefficients for any of three breaks (WTO 2001, GFC 2008, tariffs 2018). We cannot reject the null hypothesis of parameter stability. The post-GFC period is characterized by increased fiscal dominance of current account determination, with the fiscal balance coefficient strengthening, but the demographic channel remains operative.

## Coefficient Stability Across Country Groups

While the baseline coefficients are stable over time, they are sensitive to sample composition across regions. We assess this through a leave-one-region-out jackknife, dropping each of nine geographic groups in turn (SSA, Advanced Europe, MENA, Latin America, South/Southeast Asia, Central Asia/Caucasus, East Asia, Advanced Other, Other) and re-estimating the baseline model.

: Leave-One-Region-Out Jackknife: Baseline Model {#tbl:jackknife}

| Dropped Region | N countries | R-squared | Z_1 | p(Z_1) | Z_2 | p(Z_2) | Z_3 | p(Z_3) |
|:---------------|---:|---:|---:|---:|---:|---:|---:|---:|
| Adv. Europe | 108 | 0.289 | 58.7 | <0.001 | -9.4 | <0.001 | 0.40 | <0.001 |
| Adv. Other | 133 | 0.276 | 47.5 | 0.001 | -7.8 | <0.001 | 0.33 | <0.001 |
| **CCA** | **124** | **0.326** | **13.7** | **0.398** | **-2.2** | **0.351** | **0.09** | **0.334** |
| East Asia | 134 | 0.276 | 51.2 | 0.001 | -8.4 | <0.001 | 0.36 | <0.001 |
| Latin America | 122 | 0.282 | 45.1 | 0.003 | -7.6 | 0.001 | 0.33 | <0.001 |
| MENA | 122 | 0.143 | 40.4 | 0.013 | -6.9 | 0.005 | 0.30 | 0.002 |
| Other | 132 | 0.280 | 49.4 | 0.001 | -8.0 | <0.001 | 0.34 | <0.001 |
| S/SE Asia | 122 | 0.303 | 58.2 | <0.001 | -9.6 | <0.001 | 0.41 | <0.001 |
| SSA | 99 | 0.243 | 63.8 | <0.001 | -10.0 | <0.001 | 0.42 | <0.001 |
| **Full sample** | **137** | **0.273** | **48.2** | **0.001** | **-7.9** | **<0.001** | **0.34** | **<0.001** |

The coefficient of variation across jackknife drops is approximately 30 percent for all three Z terms. Critically, the Central Asia and Caucasus group (CCA: Armenia, Azerbaijan, Belarus, Georgia, Kazakhstan, Kyrgyzstan, Moldova, Mongolia, Russia, Tajikistan, Turkmenistan, Ukraine, Uzbekistan) is the only regional drop that flips all three demographic coefficients from significant to non-significant. These 13 transition economies have idiosyncratic current account dynamics---commodity export revenues, remittance dependence, and post-Soviet structural adjustment---combined with distinctive demographic profiles (rapid fertility declines from Soviet-era high levels). Their inclusion provides the variation that tips the demographic coefficients from marginal to highly significant.

A cumulative sample build-up confirms this pattern. Starting from the EBA-49 core (Z_1 = 17.6, p = 0.41), successively adding Advanced Europe, Latin America, MENA, and South/Southeast Asia leaves Z_1 at 8.6 (p = 0.61). Adding the 12 CCA countries then jumps the coefficient to 51.9 (p = 0.001)---a sixfold increase from one regional addition.

The KAOPEN interaction terms (Model 3) are even more sensitive. Dropping either SSA or South/Southeast Asia eliminates significance for all three Z $\times$ KAOPEN terms. The Z_1 $\times$ KAOPEN interaction loses significance when any of five out of nine regions is dropped.

A decomposition of the CCA group into commodity exporters (Azerbaijan, Kazakhstan, Russia, Turkmenistan, Uzbekistan) and non-commodity economies (Armenia, Belarus, Georgia, Kyrgyzstan, Moldova, Mongolia, Tajikistan, Ukraine) reveals that the tipping point is the non-commodity subgroup. Dropping the 5 commodity CCA countries has no effect on the results (Z_1 = 49.2, p = 0.001). Dropping the 8 non-commodity CCA countries eliminates significance entirely (Z_1 = 14.3, p = 0.37). These 8 countries are predominantly remittance-dependent economies with large diasporas and structural current account deficits (mean CA/GDP = -7.4%) combined with distinctive post-Soviet demographic profiles. When CCA membership is controlled for through dummy variables, the non-commodity CCA dummy is highly significant (-6.7, p < 0.001) and the Z coefficients actually strengthen (Z_1 rises from 48.2 to 59.7, p < 0.001). This suggests that the CCA non-commodity countries have a level shift in CA/GDP---driven by remittance inflows and structural factors---that is correlated with their demographic position but not caused by it. The correlation between their age structure and their current account position helps identify the Z coefficients, but the mechanism may not be the lifecycle channel.

### Remittance Robustness Test

A natural concern is that the CCA non-commodity countries drive significance not through lifecycle savings but through remittance-driven consumption patterns. We test this directly by adding World Bank personal remittances received (% of GDP) as a control variable. CCA countries average 7.9% remittances/GDP versus 3.5% for the rest of the sample, confirming high remittance dependence.

Adding remittances to the baseline model has virtually no effect on the demographic coefficients: Z_1 moves from 48.3 to 46.4, with all three Z terms retaining p < 0.002. Remittances themselves are insignificant (p = 0.21). Critically, dropping CCA countries still eliminates demographic significance regardless of whether remittances are controlled for (all Z p > 0.66 with remittances, p > 0.35 without). The remittance-driven consumption hypothesis does not explain the CCA sensitivity.

A more revealing test interacts the Z polynomials with a CCA membership indicator. The CCA×Z interaction terms are highly significant (all p < 0.01), while the non-CCA Z coefficients are individually insignificant (all p > 0.32). The CCA demographic slopes are 3-8 times larger than the rest-of-world slopes (CCA Z_1 = 109.9 vs. non-CCA Z_1 = 17.4). Three aspects of the CCA sensitivity should be distinguished. First, the statistical fact: CCA demographic slopes are significantly larger than the rest-of-world slopes, and this is a slope difference (interaction effect), not merely a level shift or remittance confound. Second, the interpretation: this could reflect genuinely stronger lifecycle dynamics in transition economies---where compressed fertility declines, rapid capital account liberalization, and structural reorganization amplify the demographic-savings nexus---or it could reflect post-Soviet structural factors that are correlated with but not caused by demographic change. The event study evidence (demographics predict CCA current accounts better before formal liberalization than after) complicates the pure lifecycle reading. Third, the implication: pooled coefficients are not transportable. The full-sample estimates reflect a weighted average of CCA and non-CCA slopes, and applying them uniformly to all countries would overstate effects for non-CCA economies. The jackknife-mean coefficient or region-specific estimates provide more conservative benchmarks.

An event study testing whether the demographic signal "activates" upon capital account opening in CCA countries yields inconclusive results: demographics predict CCA current accounts better *before* formal liberalization (all Z p < 0.04 pre-opening) than after (Z_1 p = 0.62 post-opening). This pattern is difficult to reconcile with a pure lifecycle/openness-gating story, though it may reflect the gradual nature of effective financial integration---legal opening (measured by KAOPEN) precedes the development of financial infrastructure, correspondent banking relationships, and investor trust by many years.

These findings warrant four observations. First, the statistical significance of the demographic polynomial terms depends on particular country groups whose inclusion may introduce correlated unobservables. Second, the CCA sensitivity is specifically a non-commodity phenomenon---not an oil revenue confound or a remittance confound, as direct testing confirms. Third, the CCA demographic slopes are genuinely larger than the rest-of-world slopes, suggesting heterogeneous effect sizes rather than a statistical artifact, though the mechanism remains debatable. Fourth, the coefficient magnitudes drive the projection exercise (Section 7), meaning that the projected swings should be interpreted as indicative of direction and relative magnitude rather than as precise point estimates.

### Jackknife Coefficient Ranges

For policy applications, we report the full range of coefficients across jackknife drops alongside the full-sample estimates:

: Jackknife Coefficient Ranges {#tbl:jk_ranges}

| Variable | Full Sample | JK Mean $\pm$ SD | JK Range | CV | Sig. at 5% |
|:---------|:-----------|:-----------------|:---------|:---|:-----------|
| Z_1 | 48.3 (p=0.001) | 47.6 $\pm$ 13.8 | [13.7, 63.8] | 29% | 8/9 drops |
| Z_2 | -7.9 (p<0.001) | -7.8 $\pm$ 2.2 | [-10.0, -2.2] | 28% | 8/9 drops |
| Z_3 | 0.34 (p<0.001) | 0.33 $\pm$ 0.09 | [0.09, 0.42] | 28% | 8/9 drops |
| R² | 0.273 | 0.273 $\pm$ 0.05 | [0.143, 0.326] | --- | --- |
| fiscal_bal | 0.307 | 0.309 $\pm$ 0.06 | [0.201, 0.402] | --- | --- |

The coefficients are significant at the 5% level in 8 of 9 jackknife samples, with CCA being the sole exception. The coefficient of variation is approximately 29% for all three Z terms---substantial but not catastrophic. Policy users should treat the jackknife mean as a conservative central estimate and the range as a plausible interval.


# Demographic Projections

Using the UN WPP 2024 medium-variant population projections and our estimated baseline coefficients, we project the demographic contribution to current account balances through 2060 for all 141 countries. These projections isolate the mechanical effect of changing age structures, holding all other determinants constant. Given the coefficient sensitivity documented in Section 6.3, the magnitudes should be interpreted as indicative rather than precise; the directions and relative country rankings are robust across sample specifications, but the absolute values scale linearly with the estimated Z coefficients, which exhibit a CV of approximately 30% across regional jackknife drops (Z_1 ranges from 13.7 to 63.8; Table @tbl:jk_ranges). All projections below use full-sample coefficients; conservative estimates using jackknife-mean coefficients would be approximately 2% smaller.

: Projected Demographic Contribution to CA/GDP (percentage points) {#tbl:projections}

| Country | 2025 | 2030 | 2040 | 2050 | 2060 | Swing 25-50 |
|:--------|---:|---:|---:|---:|---:|---:|
| CHN | -3.9 | -3.5 | +1.0 | +5.1 | +8.4 | +9.1 |
| IND | -2.5 | -4.2 | -4.6 | -5.6 | -5.0 | -3.1 |
| JPN | +10.5 | +11.6 | +14.1 | +16.5 | +18.7 | +6.0 |
| USA | +1.0 | +1.6 | +4.6 | +3.9 | +3.1 | +2.9 |
| DEU | +3.7 | +3.7 | +9.0 | +10.8 | +8.9 | +7.1 |
| KOR | -1.7 | -0.6 | +7.7 | +14.6 | +19.2 | +16.3 |
| BRA | -3.2 | -4.0 | -2.0 | -0.8 | +1.6 | +2.4 |
| NGA | -0.7 | -1.9 | -2.3 | -4.8 | -7.7 | -4.2 |
| IRN | -4.7 | -6.1 | -5.1 | -2.8 | +2.7 | +2.0 |
| VNM | -4.2 | -5.0 | -2.9 | -1.7 | +0.2 | +2.5 |
| BGD | -1.3 | -3.1 | -3.4 | -4.7 | -4.3 | -3.4 |
| GTM | -0.1 | -2.2 | -4.2 | -6.9 | -7.5 | -6.7 |
| SVN | +0.9 | +1.9 | +7.3 | +9.8 | +11.4 | +8.8 |
| HKG | -1.0 | +1.2 | +8.7 | +19.9 | +28.5 | +20.9 |
| TWN | -2.8 | -0.6 | +6.5 | +12.4 | +18.7 | +15.2 |
| ESP | +1.4 | +2.7 | +8.5 | +13.5 | +15.3 | +12.1 |

![Projected demographic pressure on current accounts for selected economies, 1980-2070. The dashed vertical line marks 2024 (present).](figures/fig6_projections.png){#fig:projections}

Of 141 countries with projections, 92 currently face demographic headwinds (contribution below -0.5 pp), 30 enjoy tailwinds (above +0.5 pp), and 19 are near zero. The headwind count is moderately sensitive to the threshold: 105 at -0.25 pp, 92 at -0.5 pp, and 74 at -1.0 pp. Country rankings by projected demographic contribution are robust across jackknife samples (Spearman rank correlation of top and bottom 30 exceeds 0.93 across all regional drops). The geographic pattern is clear: advanced economies and rapidly aging East Asian economies generate positive contributions, while developing economies across Sub-Saharan Africa, South Asia, and Central America face negative contributions that will deepen through 2050.

**China** represents the most dramatic reversal. The demographic contribution is currently -3.9 pp, reflecting the large working-age cohort's savings being offset by the beginning of rapid aging. By 2040, the contribution turns positive as China's elderly share surpasses the world average, reaching +5.1 pp by 2050 and +8.4 pp by 2060. This 9.1 pp swing will create powerful demographic pressure toward current account surplus---or, equivalently, capital export---in the world's second-largest economy.

**Korea** faces the largest projected swing of any major economy (+16.3 pp by 2050), driven by one of the world's lowest fertility rates. Hong Kong (+20.9 pp) and Taiwan (+15.2 pp) show even larger swings, reflecting similarly extreme aging trajectories.

**India** follows a deepening deficit trajectory (-2.5 pp in 2025 to -5.6 pp by 2050), driven by the expanding working-age share creating investment demand that exceeds domestic savings. India's demographic window for capital absorption extends through the 2050s.

**Iran** (-4.7 pp in 2025, +2.0 pp swing to 2050) illustrates a fertility-crash trajectory. The post-revolutionary decline from TFR 6.5 to 1.7 creates demographic dynamics comparable to Korea's, compressed into a shorter window. With KAOPEN at -1.25, these pressures do not currently transmit to capital flows.

**Guatemala** has the largest remaining demographic dividend in the Western Hemisphere, with a projected -6.7 pp swing to 2050---comparable to Nigeria's trajectory but compressed into a shorter window.

**Nigeria** and Sub-Saharan Africa more broadly represent the largest pool of demographically young economies. Nigeria's projected contribution deepens from -0.7 pp in 2025 to -7.7 pp by 2060 (a -4.2 pp swing), as continued high fertility sustains a large youth dependency burden relative to the aging global average. Whether this demographic pressure translates into capital inflows depends critically on financial integration: with KAOPEN at -0.68, Nigeria's capital account remains substantially closed, and our three-way interaction results (Section 5.3) show that the demographic-current account channel activates only as developing economies open. SSA thus represents both the largest future source of demographic-driven capital demand and the region where institutional barriers are most likely to prevent that demand from materializing.

**The Baltic-Balkan pattern**: Slovenia (+8.8 pp swing), Bosnia-Herzegovina (+8.5 pp), Slovakia (+6.9 pp), and others show sharp recovery trajectories as they emerge from demographic troughs following the 1990s population upheavals.

## General Equilibrium Clearing Rate Overlay

The projections above are partial equilibrium (PE): they apply historical coefficients to future demographics, holding the global interest rate fixed. We implement a GE clearing overlay following the logic of @carvalho2016 and @rachel2017. This measure represents the GDP-weighted sum of partial-equilibrium demographic contributions across countries---a diagnostic of the rate channel's potential magnitude, not an equilibrium condition literally violated, since actual world current accounts sum to zero by construction.

$$\sum_i w_i \left[ \widehat{CA}^{PE}_i + \delta \left( \widehat{\Delta yield}_i - \Delta r^* \right) \right] = 0$$

where $\delta = 0.127$ is the rate-to-CA semi-elasticity from Model 3b and $\widehat{\Delta yield}_i$ is the projected yield differential from a Stage 1 demographics-to-yield regression.

: Global Clearing Rate Adjustment {#tbl:ge_clearing}

| Year | PE Demographic Surplus Pressure (pp) | $\Delta r^*$ (pp) | % Cleared by Rates |
|:-----|-------------------------:|-------------------:|-------------------:|
| 2025 | -0.14 | -1.08 | 100% |
| 2030 | +0.01 | +0.10 | 100% |
| 2035 | +2.51 | +2.00 | 12% |
| 2040 | +3.09 | +2.00 | 10% |
| 2050 | +4.16 | +2.00 | 8% |
| 2060 | +4.98 | +2.00 | 6% |

*Note: $\Delta r^*$ is capped at $\pm$2 pp. When the cap binds (2035--2060), the residual imbalance must be absorbed through other channels.*

With near-universal country coverage, the GDP-weighted demographic surplus pressure grows substantially after 2035, reaching +4.16 pp by 2050. This reflects the asymmetric composition of the 140-country sample: aging advanced economies with large GDP weights generate positive demographic CAs, while many young developing economies with small individual GDP weights generate negative CAs that do not fully offset in aggregate. The clearing rate cap binds continuously from 2035 to 2060, and the rate channel absorbs only 6-12% of this surplus pressure---far less than the 45-67% implied by studies limited to advanced economies.

This finding has an important interpretation, though its magnitude is sensitive to the estimated coefficients. The 140-country Z coefficients are substantially larger than EBA-49 subsample estimates (Section 6.3), which amplifies all PE projections. With EBA-49 subsample coefficients (approximately 40--60% of the full-sample estimates), the PE imbalance at 2050 would be substantially smaller, and the rate channel would clear nearly all of it. The truth likely lies between these extremes. Nevertheless, two findings are robust to coefficient choice: first, the rate channel has a hard ceiling ($\delta \times 2$ pp = 0.254 pp per country), which becomes binding whenever the PE imbalance exceeds approximately 0.5 pp; and second, the asymmetric sample composition---aging advanced economies generating surpluses that small young developing economies cannot offset---ensures that the net surplus pressure grows with sample coverage regardless of coefficient values. The remaining imbalance must be absorbed through exchange rate adjustment, fiscal policy responses, capital flow management, and structural changes in savings-investment behavior.

: PE vs GE Demographic Current Account Contributions (pp of GDP) {#tbl:pe_vs_ge}

| Country | | 2030 | 2040 | 2050 | 2060 |
|:--------|:--|-----:|-----:|-----:|-----:|
| Japan | PE | +11.6 | +14.1 | +16.5 | +18.7 |
| | GE | +11.5 | +13.7 | +16.1 | +18.4 |
| Korea | PE | -0.6 | +7.7 | +14.6 | +19.2 |
| | GE | -0.7 | +7.2 | +14.1 | +18.8 |
| China | PE | -3.5 | +1.0 | +5.1 | +8.4 |
| | GE | -3.6 | +0.6 | +4.7 | +7.9 |
| USA | PE | +1.6 | +4.6 | +3.9 | +3.1 |
| | GE | +1.6 | +4.3 | +3.6 | +2.8 |
| India | PE | -4.2 | -4.6 | -5.6 | -5.0 |
| | GE | -4.1 | -4.8 | -5.9 | -5.3 |
| Nigeria | PE | -1.9 | -2.3 | -4.8 | -7.7 |
| | GE | -1.8 | -2.4 | -4.9 | -7.8 |

The GE adjustments are moderate---typically 0.2-0.5 pp per country---confirming that PE projections are reasonable first approximations. The adjustment systematically reduces projected surpluses for aging economies (Japan at 2040: 14.1 to 13.7; China at 2050: 5.1 to 4.7) and modestly deepens deficits for young economies (India at 2050: -5.6 to -5.9).


# Discussion

## The Dual-Channel Finding

The paper identifies a dual institutional channel through which demographics affect current accounts. In developing countries, financial openness gates the demographic-current account mechanism: closed economies with young populations show no demographic effect, while similar economies that have opened begin to exhibit the lifecycle pattern. A horse race against income-level interactions confirms this is genuinely about financial openness (KAOPEN interactions survive at p < 0.01 while GDP interactions collapse to p > 0.49). In advanced economies, pension system generosity determines the strength of the demographic-CA relationship.

This dual-channel result has two implications. First, it reconciles the apparent tension between @higgins1998 (who emphasized financial openness) and @boersch2010 (who emphasized pension systems) by showing that both are correct but in different country groups. Second, it generates distinct policy recommendations for developing vs. advanced economies (see below).

However, the coefficient instability documented in Section 6.3 tempers these findings. The demographic coefficients that drive the dual-channel analysis are themselves sensitive to sample composition, particularly the inclusion of CCA transition economies. The dual-channel structure---openness for developing countries, pensions for advanced economies---is robust to sample variation (the three-way interaction result holds across jackknife drops), but the magnitudes of the underlying demographic effects are not. Future work should investigate whether the CCA countries' contribution reflects genuine demographic signal amplified by their extreme demographic transitions, or whether it reflects omitted variables specific to post-Soviet economies.

What should practitioners take from this fragility? Three operational conclusions follow. First, the pooled average demographic effect exists but is not policy-transportable: applying the full-sample coefficients uniformly to all countries would overstate effects for non-CCA economies. Region-specific slopes or the jackknife-mean coefficient (approximately 2% below the full-sample estimate) provide more conservative benchmarks. Second, the EBA framework should treat demographics as a medium-term directional anchor---the *signs* and *relative country rankings* of projected effects are robust across all sample specifications, even when magnitudes vary by a CV of 30%---rather than as a source of precise point estimates. Third, the dual-channel institutional structure (openness for developing, pensions for advanced) is the most stable finding and should receive greater weight than the baseline polynomial coefficients themselves.

## Financial Openness and the Lucas Paradox

Our finding that the KAOPEN interaction operates exclusively through developing countries provides a demographic lens on the Lucas paradox. Capital does not flow from rich to poor countries partly because many poor countries' capital accounts are closed, preventing demographic forces from transmitting internationally. As developing economies open---and our data show that this process is ongoing, with median KAOPEN rising from -0.5 in 1990 to -0.06 by 2020---the lifecycle pattern should increasingly manifest in their current accounts, potentially unlocking capital flows to young, capital-scarce economies.

## Residual Analysis

: Country Residuals from Baseline Model (Top and Bottom 10) {#tbl:residuals}

![](tables/country_residuals.md)

Countries where actual current accounts systematically deviate from model predictions include Singapore, Switzerland, and Nigeria (largest positive residuals), attributable to financial center effects, safe-haven capital inflows, and oil export revenues respectively. The persistent US negative residual reflects the dollar's reserve currency role, which generates capital inflows beyond what demographics and macro fundamentals predict.

## Limitations

Several limitations warrant acknowledgment. First, and most importantly, the leave-one-region-out jackknife (Section 6.3) reveals that the baseline demographic coefficients depend critically on the inclusion of 13 Central Asia and Caucasus transition economies. These countries share post-Soviet structural features (commodity dependence, migration corridors, rapid institutional change) that may confound the demographic channel. The coefficient of variation across regional drops is approximately 30 percent, and the KAOPEN interaction terms are even more sensitive, losing significance when any of five out of nine regions is dropped. The standard EBA-49 subsample produces insignificant Z coefficients (Z_1 p = 0.41), confirming that the broader literature's difficulty in establishing robust demographic-current account coefficients reflects genuine fragility rather than insufficient data. Second, while the interest rate channel is significant when measured through government bond yields, the bond yield sample is restricted to 23 OECD economies, and the two-stage Carvalho test produces only marginal significance in Stage 1 (individual Z p-values of 0.12--0.18). Third, the pension model relies on OECD SOCX data covering only 40 countries; broader pension coverage data from the World Bank ASPIRE database (118 countries) would provide a more definitive test. Fourth, the post-2019 estimation sample is thin due to data availability lags in several sources (PWT, KAOPEN, WDI). Fifth, the pooled GLS approach, while consistent with EBA practice, does not capture time-invariant country heterogeneity that fixed effects would address; however, Appendix E shows that country fixed effects produce *larger* demographic coefficients (all p < 0.001), and Appendix F confirms the results in long differences. Sixth, the three-way interaction analysis (Section 5.3) is demanding of the data, with only 21 high-income countries in the extended model; confirmation with alternative income classifications or continuous interaction specifications would strengthen the finding.

## Policy Implications

Five policy implications emerge. First, developing countries with favorable demographic trajectories should pursue financial integration to realize the capital flow effects of their changing age structures---our three-way interaction results show that this channel is active specifically for them. Second, advanced economies facing rapid aging should attend to pension system design as the primary institutional channel through which demographics affect external balances; pension reform will alter how aging translates to current account positions. Third, the near-universal GE clearing analysis shows that the interest rate channel is far weaker as an equilibrating mechanism than AE-focused studies suggest, implying that exchange rate flexibility and fiscal adjustment must bear most of the burden. Fourth, the 92 countries currently facing demographic headwinds on their current accounts span all income levels and regions, underscoring that demographic transition management is a global policy challenge, not one confined to advanced economies. Fifth, the absence of statistically significant structural breaks in the demographic-current account relationship, when estimated on the broad sample, suggests that demographics provide a directional medium-term anchor for current account assessment---a finding directly relevant to the IMF's External Balance Assessment methodology.


# Conclusion

This paper provides a comprehensive investigation of the relationship between demographic structure and international capital flows across 140 countries. Using the Fair-Dominguez polynomial technique embedded in the IMF's EBA framework, we find that all three demographic polynomial variables are individually highly significant (all p < 0.001) in the full sample---a result enabled by the breadth of our 140-country sample. The broad sample resolves several apparent puzzles in the existing literature: the post-GFC weakening of demographic effects was a sample composition artifact; the financial openness interaction is a developing-country phenomenon rather than universal; and pension system generosity is the relevant institutional mediator for advanced economies.

However, a thorough stability analysis reveals that these results are more fragile than the headline significance suggests. A leave-one-region-out jackknife shows that the demographic coefficients depend critically on the inclusion of 13 Central Asia and Caucasus transition economies, with a coefficient of variation of approximately 30 percent across regional drops. The CCA countries exhibit demographic slopes 3--8 times larger than the rest of the world, and this amplified sensitivity---rather than statistical contamination---drives the full-sample significance. Direct testing rules out remittance dependence as the mechanism (remittances/GDP is insignificant at p = 0.21 and barely changes the Z coefficients), but an event study finds that demographics predict CCA current accounts better before formal capital account opening than after, complicating the lifecycle interpretation. The KAOPEN interaction terms are even more sensitive, losing significance when any of several regions is dropped. These findings suggest that the broader literature's difficulty in establishing robust demographic-current account effects reflects genuine identification challenges: the effect appears to operate with heterogeneous intensity across country groups, and universal pooled estimates obscure this heterogeneity.

The key contributions are: first, assembling the broadest country sample in the demographic-current account literature and documenting how results depend on sample composition; second, identifying a dual institutional channel where financial openness matters for developing countries and pension systems for advanced economies, confirmed through a horse race that distinguishes openness from income level; third, documenting that the interest rate channel appears more clearly through bond yields and term spreads, though the demographics-to-yields link remains suggestive; and fourth, providing near-universal projections with a GE clearing overlay that reveals the rate channel's limited capacity to equilibrate global demographic imbalances.

Looking forward, the projections highlight a geographic rotation of demographic pressure, with 92 countries currently facing headwinds and the largest adjustments ahead for East Asia. The magnitudes of projected effects should be interpreted cautiously given the coefficient sensitivity documented above, but the directions and relative rankings are robust across sample specifications. The general equilibrium analysis shows that with near-universal coverage, interest rate adjustment can absorb only 6--12 percent of peak demographic surplus pressure, placing the primary equilibrating burden on exchange rates, fiscal policy, and structural reform. Whether these demographic forces manifest in actual capital movements depends on the institutional environment---financial openness for developing economies, pension design for advanced ones---a finding that connects the academic literature to policy design while underscoring the need for further work on the stability of demographic-current account coefficients across country samples.


# References

::: {#refs}
:::


\newpage

# Appendix

## A. Variable Definitions

| Variable | Definition | Source |
|:---------|:-----------|:-------|
| CA/GDP | Current account balance as percentage of GDP | IMF WEO |
| $Z_1, Z_2, Z_3$ | Demographic polynomial variables from GDP-weighted demeaned age shares | UN WPP 2024 |
| Fiscal balance/GDP | General government net lending/borrowing as percentage of GDP | IMF WEO |
| KAOPEN | Chinn-Ito financial openness index (normalized) | Chinn and Ito |
| Expected growth | 5-year-ahead IMF WEO GDP growth forecast | IMF WEO |
| NFA/GDP (lagged) | Previous year's net foreign assets as percentage of GDP | Lane and Milesi-Ferretti EWN |
| Log relative OPW | Log of output per worker relative to US, PPP-adjusted | Penn World Tables 10.0 |
| Life expectancy | Life expectancy at birth (years) | World Bank WDI |
| Health exp./GDP | Government health expenditure as percentage of GDP | World Bank WDI |
| Log lending rate | log(1 + lending_rate/100); continuously compounded domestic lending rate | IMF MFS, FRED |
| Pension spending/GDP | Public expenditure on old-age and survivors as percentage of GDP | OECD SOCX |

## B. Country Sample

**EBA 49**: ARE, ARG, AUS, AUT, BEL, BRA, CAN, CHE, CHL, CHN, COL, CZE, DEU, DNK, EGY, ESP, FIN, FRA, GBR, GRC, HKG, HUN, IDN, IND, IRL, ISR, ITA, JPN, KOR, MAR, MEX, MYS, NLD, NOR, NZL, PAK, PER, PHL, POL, PRT, RUS, SAU, SGP, SWE, THA, TUR, TWN, USA, ZAF

**Sub-Saharan Africa (20)**: AGO, BFA, BWA, CIV, CMR, ETH, GHA, KEN, MDG, MOZ, MUS, MWI, NAM, NGA, RWA, SEN, TZA, UGA, ZMB, ZWE

**EU Completion (10)**: BGR, CYP, EST, HRV, LTU, LVA, MLT, ROU, SVK, SVN

**Additional Economies (61)**: ALB, ARM, AZE, BDI, BEN, BGD, BHR, BIH, BLR, BOL, BTN, CAF, COM, CPV, CRI, DOM, DZA, ECU, GAB, GEO, GIN, GNB, GNQ, GTM, HND, IRN, IRQ, JAM, JOR, KAZ, KGZ, KHM, KWT, LAO, LBN, LBR, LSO, MDA, MKD, MLI, MMR, MNG, NPL, OMN, PRY, QAT, SDN, SLE, SWZ, SYC, TCD, TGO, TJK, TKM, TUN, UKR, URY, UZB, VEN, VNM, YEM

## C. Model Comparison

![Model comparison: R-squared across four specifications. Observation counts shown in parentheses.](figures/fig7_model_comparison.png){#fig:modelcomp}

## D. Subsample Polynomial Re-Estimation

As a robustness exercise, we re-estimate the baseline model on subsamples defined by geography, income, and institutional membership. The Fair-Dominguez polynomial has only three free parameters ($\gamma_1$, $\gamma_2$, $\gamma_3$) constraining the shape of 17 age-group coefficients. In subsamples with fewer than approximately 25 countries, particularly those with similar demographic structures, there is insufficient cross-country variation in age distributions to identify the lifecycle profile. The main-body analysis uses full-sample projected contributions (Table 5) to avoid these small-sample instabilities while preserving the heterogeneity narrative through formal interaction tests (Table 6).

## E. Country Fixed Effects Robustness

The pooled GLS specification used throughout the paper does not control for time-invariant country heterogeneity. A referee concern is that slow-moving demographic variables may be identified primarily from cross-sectional variation that is confounded with time-invariant institutional or structural factors. We address this by estimating country-and-year fixed effects models that identify the demographic coefficients exclusively from within-country variation over time.

**Table E1: Country + Year Fixed Effects vs. Pooled GLS**

| | FE: Demographics only | FE: Full model | Pooled GLS |
|:--|:--:|:--:|:--:|
| $Z_1$ | 11.3 (0.130) | **62.4** (<0.001) | 50.3 (<0.001) |
| $Z_2$ | $-2.0$ (0.060) | **$-10.4$** (<0.001) | $-8.1$ (<0.001) |
| $Z_3$ | 0.10 (0.026) | **0.46** (<0.001) | 0.34 (<0.001) |
| fiscal_bal_gdp | --- | 0.48 (<0.001) | 0.32 (<0.001) |
| kaopen | --- | $-0.23$ (0.291) | $-0.80$ (<0.001) |
| expected_growth | --- | $-0.44$ (<0.001) | $-0.24$ (<0.001) |
| nfa_gdp_lag | --- | 0.11 (0.536) | 0.62 (0.004) |
| log_rel_opw | --- | 3.24 (<0.001) | 3.83 (<0.001) |
| health_exp_gdp | --- | $-0.40$ (0.092) | $-0.68$ (0.003) |
| R² (within) | 0.014 | 0.189 | 0.271 |
| N | 7,706 | 2,746 | 2,746 |
| Countries | 193 | 138 | 138 |

*Notes:* FE models include country and year fixed effects. Pooled GLS uses Prais-Winsten AR(1) correction. P-values in parentheses.

Without controls, the demographic variables are only marginally significant under country FE, reflecting the difficulty of identifying slowly evolving polynomial variables from within-country variation alone (within R² = 0.014). However, in the full model with controls, all three Z terms are highly significant (all p < 0.001) and the coefficients are *larger* than the pooled GLS estimates (Z₁: 62.4 vs. 50.3). This strengthening occurs because country FE absorb time-invariant confounders that partially offset the demographic effect in pooled estimation. The within R² of 0.189 indicates that demographics and controls explain 19% of within-country current account variation after removing permanent country-level differences.

Two variables lose significance under country FE: KAOPEN (p = 0.29) and NFA/GDP (p = 0.54). Both are persistent within-country series whose identification relies substantially on cross-sectional variation. The remaining controls retain their signs and significance, with fiscal balance and expected growth becoming *stronger* under FE---consistent with these variables having meaningful within-country time variation.

The meta-point across Appendices E--G is that identification of demographic effects on current accounts is inherently difficult because demographics evolve slowly. The results survive multiple approaches (country FE, long differences, alternative SE corrections) but remain heterogeneous in magnitude. Researchers should interpret this not as evidence against demographic effects, but as a feature of any slowly evolving regressor in country panel data: the effect is real, but precisely estimating its magnitude requires accepting either cross-sectional identification (pooled GLS, larger N) or within-country identification (FE/long differences, larger coefficients but wider confidence intervals).

## F. Long Differences

To further address concerns about serial correlation driving inference in the pooled panel, we estimate long-difference specifications that regress 10-year and 5-year changes in CA/GDP on corresponding changes in all regressors. Long differences absorb both country-specific levels and slow trends, identifying the demographic effect from medium-run within-country variation.

**Table F1: Long-Difference Estimates**

| | 10-year $\Delta$ | 5-year $\Delta$ |
|:--|:--:|:--:|
| $\Delta Z_1$ | 27.1 (0.022) | 43.9 (0.003) |
| $\Delta Z_2$ | $-5.5$ (0.001) | $-7.6$ (<0.001) |
| $\Delta Z_3$ | 0.28 (<0.001) | 0.34 (<0.001) |
| $\Delta$ fiscal_bal_gdp | 0.42 (<0.001) | 0.49 (<0.001) |
| $\Delta$ kaopen | 0.11 (0.703) | $-0.72$ (0.007) |
| $\Delta$ expected_growth | $-0.15$ (0.006) | $-0.38$ (<0.001) |
| R² | 0.189 | 0.167 |
| N | 1,307 | 1,983 |

*Notes:* OLS on non-overlapping 10-year (5-year) differences. P-values in parentheses.

All three demographic terms survive long differencing at both horizons. The 10-year specification (N = 1,307 non-overlapping differences) produces somewhat attenuated Z₁ (27.1 vs. 50.3 pooled) but highly significant Z₂ and Z₃. The 5-year specification (N = 1,983) produces coefficients closer to the pooled estimates. These results confirm that within-country medium-run changes in demographic structure predict medium-run changes in current account balances conditional on changes in macro controls, substantially reducing reliance on cross-sectional identification. We note, however, that long differencing does not fully eliminate joint-trend concerns: demographics trend and current accounts have persistent components, so coincident medium-run shifts could still contribute. The inclusion of differenced controls (fiscal balance, expected growth, KAOPEN) absorbs some of this trending, and the attenuation of Z₁ from 50.3 to 27.1 in 10-year differences is consistent with removing some cross-sectional confounding rather than merely reflecting measurement noise.

## G. Country-Clustered Standard Errors

We also report country-clustered standard errors (Arellano, 1987) to assess whether inference depends on the AR(1) parametric correction. Clustering by country allows arbitrary within-country error dependence.

The demographic polynomial terms become individually insignificant under country-clustered SEs (Z₁ p = 0.16, Z₂ p = 0.17, Z₃ p = 0.19), with standard errors approximately 1.5--2.4 times the GLS SEs. This reflects the conservative nature of clustering with only 138 clusters on slowly evolving regressors. Country-clustered SEs are known to be severely biased in panels where $T$ is large relative to the number of clusters and regressors are persistent [@bertrand2004]. Our pooled GLS AR(1) correction is the appropriate remedy for serial correlation in this setting, but we report clustered SEs for transparency. The long-difference results (Appendix F), which are immune to serial correlation concerns, confirm that the demographic effects are genuine.

## H. CCA Influence Diagnostics

Section 6.3 documents that the CCA region is the sensitivity tipping point for demographic coefficients. To understand whether particular CCA countries drive this, we compute leave-one-country-out DFBETA influence statistics for $Z_1$.

The most influential country is Mongolia (DFBETA = +29.7, a CCA member), whose removal reduces the Z₁ coefficient from 50.3 to 20.6. Among the 20 most influential countries by absolute DFBETA, only 5 are CCA members (Mongolia, Kyrgyzstan, Azerbaijan, Tajikistan, Uzbekistan); the remaining 15 include Bhutan ($-9.2$), Nigeria ($-3.5$), Algeria (+3.5), Greece ($-2.5$), and Kuwait ($-2.5$). The CCA influence is thus concentrated in Mongolia and distributed across a small number of other CCA countries, rather than being a uniform region-wide effect.

We also compare CCA vs. non-CCA observable characteristics. CCA countries have significantly different demographics (Z₁: $-1.19$ vs. $-1.49$, p < 0.001), lower NFA positions ($-0.61$ vs. $-0.25$, p < 0.001), lower KAOPEN ($-0.09$ vs. +0.13, p = 0.001), and higher trade openness (94.5% vs. 81.5%, p < 0.001). These differences are consistent with the post-Soviet transition features discussed in Section 6.3 and suggest that the CCA demographic slopes may reflect genuine heterogeneity in how demographic forces translate to external balances in transition economies.

## I. Interest Rate Data Coverage

A referee concern is that our interest rate channel tests (Section 5.6) are limited to OECD economies. We report the coverage matrix for all rate variables:

| Variable | N obs | Countries | Year range |
|:---------|------:|----------:|:-----------|
| Government bond yield (10yr) | 1,018 | 23 | 1970--2024 |
| Short-term rate (3m) | 983 | 23 | 1970--2024 |
| Policy rate | 2,071 | 81 | 1970--2024 |
| Lending rate | 4,297 | 132 | 1970--2024 |
| Term spread | 931 | 23 | 1970--2024 |

The bond yield sample (23 countries) represents 73% of world GDP but only 16% of the 140-country sample. The lending rate covers 132 countries but, as documented in Section 5.6, contains extreme values for emerging markets (raw range up to 99,765%) that prevent meaningful cross-country comparison even after log transformation. Policy rates (81 countries) offer an intermediate option but are heterogeneous in definition across central banks. The restriction to OECD bond yields for the rate channel tests is thus a data limitation rather than a methodological choice, and the results should be interpreted as establishing that the rate channel operates in advanced economies without ruling out or confirming its operation in the broader sample.
